use OSCAR as the 'web' corpus, update tests

[rspeer]

Some things updated here:

• The source for 'web' text is now the OSCAR corpus, a cleaned and deduplicated version of Common Crawl. OSCAR replaces our previous Web sources (Leeds Web Corpus and Mokk), which were from the '00s.
• Text is tokenized using wordfreq 2.4, which changes the tokenization of apostrophes next to some vowels, particularly in French -- this change is already merged in wordfreq but it affects the test data
• Changed the lists of languages by re-evaluating some Wikipedias on the borderline, excluding languages in corpora that a recent study said were identified with less than 80% accuracy, and along those lines excluding our own Swahili Twitter data that was mostly mis-identified by FTLID

The Snakefile now prints a table of languages and how many sources they have, helping me check the level of support for a given language at a glance.

[frankier]

This also seems to address a problem where in the current version bad language tags are passed for cleaning up opus counts by introducing the new recount_messy_opus_tokens rule.

[frankier]

datasets should also be added to setup.py

[frankier]

datasets just added streaming datasets which is useful for OSCAR if you have limited disk. See this commit in my fork: frankier@6fbf8cd